Talk by Song-Chun Zhu of UCLA. Given to the Redwood Center for Theoretical Neuroscience at UC Berkeley on November 12, 2009.
Images and video are very high dimensional signals that reside in a wide spectrum of manifolds/subspaces of varying dimensions. In this talk, I will discuss two types of pure manifolds:
(i) implicit manifolds for high entropy patterns, like textures, modeled by Markov random fields, and
(ii) explicit manifolds for low entropy patterns, like textons and primitives, modeled by sparse coding.
I will show that these manifolds are connected through scaling (zooming), and present a unifying theory for learning probabilistic models by manifold pursuit through information projection. Then I will discuss how these manifolds are mixed to form middle entropy patterns, such as object templates, and integrated to generate a primal sketch representation for generic images as conjectured by David Marr in his influential book.
I will also show ongoing work on video primal sketch which integrates trackable motion, intrackable motion, and human actions.